During neural machine translation, why might a model using beam search produce grammatically incorrect output despite high BLEU scores on the training data?

💡 Explanation

The model might prioritize short, precise n-grams reflected in the BLEU score, rather than capturing long-range dependencies needed for grammatical correctness, because the BLEU score primarily assesses n-gram precision. Therefore, the model produces grammatically incorrect output, rather than focusing on semantic correctness or fluency, when maximizing BLEU score.

• Why does aphasia selectively impair a patient's ability to retrieve verbs compared to nouns when describing actions performed with tools?
• Why does coarticulation in speech production blur phonetic boundaries more for vowels than for consonants?
• Why does a mobile network channel experience reduced throughput when the signal-to-noise ratio decreases?
• Why does increased syntactic ambiguity slow reading times during sentence comprehension?
• A screenplay exhibits a sudden, jarring shift in genre mid-story; why does the audience experience cognitive dissonance?
• Why does a Mandarin Chinese speaker often perceive a foreign accent when a non-native speaker consistently pronounces each word with its citation tone in a multi-syllable phrase?